基于联邦边缘学习的梯度量化和带宽分配优化策略（英文）-豆柴文库

您所在位置：网站首页 / 基于联邦边缘学习的梯度量化和带宽分配优化策略（英文）.docx / 文档详情

在线预览结束，喜欢就下载吧，查找使用更方便

5 金币

下载文档

/ 2

下载提示

如果您无法下载资料，请参考说明：

1、部分资料下载需要金币，请确保您的账户上有足够的金币

2、已购买过的文档，再次下载不重复扣费

3、资料包下载后请先用软件解压，在使用对应软件打开

文本预览

基于联邦边缘学习的梯度量化和带宽分配优化策略（英文）
Title:GradientQuantizationandBandwidthAllocationOptimizationStrategiesforFederatedEdgeLearning
Abstract:
Federatededgelearninghasemergedasapromisingapproachtoenableefficientandprivacy-preservingmachinelearningacrossdistributededgedevices.However,thelimitedcomputationalandcommunicationresourcesofthesedevicesposechallengestotheperformanceandscalabilityoffederatedlearningsystems.Inthispaper,weproposeanovelframeworkthatcombinesgradientquantizationandbandwidthallocationoptimizationstrategiesforfederatededgelearning.Byquantizingthegradientssharedamongedgedevices,weeffectivelyreducethecommunicationoverheadandimprovethetrainingefficiency.Additionally,weoptimizethebandwidthallocationsamongdevicestofurtherenhancetheoverallsystemperformance.Experimentalresultsdemonstratetheeffectivenessandefficiencyofourproposedstrategies,pavingthewayforscalableandresource-efficientfederatededgelearning.
1.Introduction
Federatededgelearninghasgainedsignificantattentionduetoitsabilitytoleveragethevastamountofdatacollectedattheedgedeviceswhileensuringdataprivacy.Inthisdistributedlearningparadigm,edgedevicescollaborativelytrainaglobalmodelbyexchangingmodelupdatesorgradientswithacentralserver.However,thelargenumberofdevicesandthelimitedresourcesontheedgeintroducechallengessuchashighcommunicationoverheadanddegradedperformance.Thispaperproposesanovelframeworktoaddressthesechallengesbycombininggradientquantizationandbandwidthallocationoptimizationstrategiesforfederatededgelearning.
2.GradientQuantizationforCommunicationEfficiency
Toreducethecommunicationoverheadinfederatededgelearning,weadoptgradientquantizationtechniques.Insteadoftransmittingthefull-precisiongradients,edgedevicesquantizethegradientsintoalowerbitrepresentationbeforesharingthemwiththecentralserver.Thisreducesboththecommunicationlatencyandtherequirednetworkbandwidth.Weexploredifferentquantizationmethods,suchasuniformquantizationandnon-uniformquantization,tofindtheoptimalbalancebetweencommunicationefficiencyan